Agenda Machine Learning Week Europe 2023

Berlin - 15 - 16 November 2023

In this Keynote Session we delve into the powerful impact of AI technologies like ChatGPT on businesses. Understand how they are rewriting the rules, instigating innovation and causing industrial disruption. We’ll unpack the applications of generative AI, from overhauling customer support to personalizing marketing and driving data-informed decision-making. Join us as we demystify the transformative power of AI and its potential to redefine your business trajectory.

The keynote introduces Object-Centric Process Mining (OCPM) which can be seen as a major breakthrough in process mining. Traditional approaches for process modeling and process analysis tend to focus on one type of objects and each event refers to precisely one such object. OCPM takes a more holistic and comprehensive approach to process analysis and improvement by considering multiple object types and events that involve any number of objects. The keynote presents the basic concepts and the need for object-centric process mining techniques using many examples. OCPM is rapidly being adopted in commercial systems, showing its practical relevance. Process mining experts expect that OCPM will become the “normal” way of doing process mining and therefore of value to any organization that wants to improve its processes.

Foundation models are so versatile and powerful. Almost daily, additional applications of these models are discovered. Foundation models understand and generate many languages and they seem to have an immensely broad knowledge covering nearly every aspect of life. Thus, at least for large language models, one foundation model such as GPT-x should at first glance suffice, which would then be incrementally extended and improved as we have already witnessed it.

Hans will describe strong reasons for developing or expecting a variety of foundation models. To this end, Hans will highlight emerging differences among models and also discuss differences among the desiderata and expectations concerning future models. Finally, existing limitations of the models will be investigated, some of which reduce their reliability while in contrast others contribute to their overall safety.

Insights into our data science solution based on KNIME: On average, more than one in four students leaves higher education without a degree. This not only has economic consequences for the education companies, but is also a waste of resources and time for those beeing affected. The goal should therefore be to focus as an educational institution on retaining students, moreover making the study and learning environment as attractive as possible. Here, machine learning can provide significant churn factors and predictions in order to gain awareness and help educational institutions to implement more targeted approaches to support student retention and reduce churn.

This session presents a bayesian approach to model cohort-level retention rates and revenue over time. Wolt uses bayesian additive regression trees to model the retention component which they couple with a linear model to model the revenue component. This method is flexible enough to allow adding additional covariates to both model components. This bayesian model allows Wolt to quantify the uncertainty in the estimation, understand the effect of the covariates and forecast the future revenue, and retention rates. The source code is open sourced on GitHub and for the presentation Juan will use synthetic data.

Optimizing production across factories requires industrial-grade AI. Discover how Federated Learning helps tackling manufacturing challenges while retaining control of IP. Explore the case study using Katulu’s FL Suite in Siemens’ industrial PCB production to improve quality control and reduce misclassification errors. We will address key challenges, such as data onboarding and management, model lifecycle and security that are essential for the success.

This talk discusses the challenges faced by the UK Local Planning Authorities (LPAs) in managing written representations from the community for their Local Plans. The volume of information can be overwhelming, making it difficult for LPAs and the Planning Inspectorate to review and assess plans. The talk presents a case study that explores the potential of Large Language Models (LLMs) to streamline the analysis of representations, with significantly reduced processing time and improved accuracy.

Targeting the right audience is central to successful online marketing. However, traditional audience targeting (e.g. age and gender) is severely limited. Modern audience targeting incorporates and predicts real user behavior, and uses machine learning to create audience that are truly relevant for online advertising. Christoph will discuss the data sources and machine-learning methods to construct relevant audiences for online advertisers, and the tools to explore and visualize such audiences.

Industrial paper manufacturing is a highly technological process that requires continuous effort to prevent interruptions. Early detection and proper identification of problems are crucial for timely servicing, optimal personnel allocation, and avoiding breakdowns. Mondi Group and d-fine developed a smart condition monitoring solution that uses the existing technical infrastructure as a maintenance assistant

In the era of data-driven decision-making, companies are investing heavily in data platforms to structure and harness their domain knowledge. However, a challenge arises when working with Large Language Models (LLMs), as they are capable of processing only a few thousand words at once. This talk aims to explore the delicate balance between big data and small context, discussing how to effectively couple existing data platforms with the power of LLMs while considering their limitations.

Personalization is a crucial principle in content moderation and information selection for many websites, especially in e-commerce where showing relevant products to customers is key to providing a valuable service. Torge & Carsten will provide a practical guide to build a cloud-based near real-time recommender service. You will come away with a comprehensive understanding of the business considerations and technical challenges involved in building a production-ready recommender system on AWS.

Renewable energy dependence on fluctuating weather conditions may prevent a fast energy transition. Since 2013, Edison have developed custom algorithms for wind and solar forecasting, moving from proprietary tools to open-source libraries, from on-premises architecture to MLOps cloud infrastructure. This case study is aimed at data scientists and MLOps engineers willing to deal with a huge number of models. You will learn how ML/AI can boost renewable energy penetration in the market.

Large language models have taken the world by storm, but what if you’d like to use one with your internal text data? LLMs aren’t trained on your data, and finetuning a bespoke LLM is daunting. However, retrieval-augmented generation (RAG) combines LLMs with document search to create powerful generative NLP systems for your text data. In this talk we will discuss how to build effective RAG pipelines, as well as how to address potential concerns like privacy and hallucination effects.

This case study offers insights on how to maximize cross sales using a data-driven, predictive approach. The project at DIE ZEIT aimed to raise sales by tailoring offers to subscribers’ product affinities. DIE ZEIT used machine learning models based on sociodemographic data, e-mail behavior, web tracking and subscription history. Results show that leveraging subscribers’ affinities can generate considerable uplift in cross sales. A key aspect for discussion is balancing product affinities and discrete marketing objectives.

Causal Discovery goes beyond Predictive Maintenance by uncovering why a machine fails. We report on the challenge of identifying confounders in complex data & how to address that in practice. We present project results e.g., where a DiscoveryBot constantly monitors a cylinder head production line & alerts emerging causes for quality defects. We will connect to an AWS instance hosting data from that production, run causal discovery algorithms live & show how to explore the results interactively.

We present a case study examining the potential of Large Language Models like ChatGPT to transform the market research industry. Our prototype agent takes a market research interest as input, autonomously consults online sources to create a relevant questionnaire, conducts the survey using online APIs for a specified sample size and demographic, and generates a report based on the survey results. This includes quantitative data, as well as analyses of open-ended responses.

70% of AI projects fail. Why? Could it be that the “people” aspect is being ignored, including people creating AI products for internal customers? Why then not leverage the marketing framework of 5 Ps of marketing to address this challenge and learn how to market data & AI solutions? This keynote will show how this concept is being applied at a pharma company to internally market data products, addressing the product, price, place, promotion, and people … all of those elements essential for successful use of data & AI in any organization.

To improve adoption of their virtual sensors, which offer predictive maintenance analytics for the car industry, Compredict guarantees customers its machine learning-based sensors will meet a performance threshold. This guarantee is backed by Munich Re’s insurance for AI, securing customers’ ROI on their AI investment for their vehicle fleet. Alexandra and Daniel discuss why insuring your AI is a unique, powerful market signal to trust your offering, especially in industries typically hesitant to dive into new technologies.

Join us for a captivating talk on the journey of scaling ads ranking at Pinterest using innovative machine learning technologies and platforms. This presentation will showcase the transition from traditional logistic regressions to deep learning-based transformer models, incorporating sequential signals, multi-task learning, transfer learning, and GPU serving. Throughout the process, we encountered various challenges and gained valuable lessons. Discover the hurdles we overcame and the insights we gained in this talk, as we share the transformation of ads ranking at Pinterest and the lessons learned along the way.

Wrap up your first day of the conference with relaxed networking. Enjoy delicious food and drinks, engaging conversations and discussions as well as relaxed networking opportunities. There’s something for everyone!

Where? Koko & Lores; Revaler Str. 29, 10245 Berlin
When? 7.30 – 9.30pm

Shuttle buses run from the conference hotel to Koko & Lores. The first departure is at 7:00 pm, and the last one is at 7:15 pm. Please note that there is no shuttle bus back. Alternatively, you can use public transportation to reach the venue. The nearest stop is Berlin Ostkreuz. Direct connection and only 3 stops away from the S-Bahn stop Landsberger Str. close tot he conference hotel.

Data Mesh has grown beyond theory and buzzword. The initial concept as introduced by Zhamak Dehghani in 2019 is described as a socio-technological approach. It has implications on how data is organized and worked with. This keynote will reflect on a Data Mesh implementation journey, on pitfalls and how they can be avoided. In addition to the organizational transformation, Norbert will in particular describe the foundational data platform development, because it’s one of the key success factors.

Understanding how to evaluate the performance of a model is fundamental to the work of Decision Science professionals. Often, metrics used to evaluate model performance are deployed based on theory and that might not, necessarily, holistically explain the models applied benefit, or lack thereof, to the business. This session will evaluate the trade off between model performance metrics and making business impact. What is good enough?

Michael Feindt will report from his journey starting from science via startup (Blue Yonder) and growth to an international corporation and exit as well as the actual second startup. Topics will be: AI hype & real value, “new“ vs “old“ AI, gap between research & real life application, differences between startups & enterprises, tiring fight against human biases and lack of mathematical thinking, trials for explainability, importance of trust, and difference between correlation & causality.

Let’s explore the advancements in ML monitoring, from early ML 1.0 to modern LLMs. We’ll emphasize solutions to the “hallucination problem” in LLMs. Then we’ll dive into a fan favorite – where we’ll focus on RAGs, and their benefits vs. challenges. This talk offers insights into the evolving strategies for maintaining robustness and accuracy in ML models and evaluating LLMs in production – particularly valuable for ML practitioners and researchers.

At PAYBACK, a global loyalty program, a crucial aspect of upskilling initiatives within the ML/DL domain is the empowerment of citizen data scientists – subject matter experts – not typically rooted in conventional data science. Recognizing their unique position to bridge the gap between data analytics and real-world business cases, PAYBACK is leveraging AutoML tools to enhance their understanding of the ML lifecycle. Here, we present the approach adopted for training citizen data scientists.

Sustainable agriculture is crucial for our environment and food security. Partnering with GREGOIRE, a leading agriculture equipment provider, we’ve developed a cloud-provider independent and scalable solution through the AI-SPRINT project. Our approach tackles challenges such as optimized pesticide dispersion using foliage volume estimation, proactive disease detection, and yield prediction. We want to share our journey, the hurdles overcome, and the diverse AI techniques deployed.

Images are the heart of our business – but sometimes it is hard for editorial, marketing and other teams to find just the right frame or sequence based on specific search criteria. Therefore we developed a tool to help: SmartArt. SmartArt extracts a video’s metadata by using computer vision and S2T algorithms. As a result, we can solve different use cases, from finding aesthetic frames for thumbnails, over content moderation, to generating marketing material for promotion.

Conformal prediction, once a small research niche, is now drawing attention from the machine learning main stream, e.g. Amazon’s implementation within the new Fortuna library. Conformal prediction adds uncertainty quantification to any machine learning model, the confidence regions it produces are provably reliable, and it does not need unrealistically strong assumptions. Applications to classification, time series, and causal machine learning will be shown in a hands-on manner.

Dialysis machines and other MedTech equipment need to operate continuously to guarantee that patients receive proper treatment. Franziskos and his team developed a predictive maintenance solution for the dialysis machines of B. Braun. Franziskos will discuss how they went from proof of concept to global rollout in a big data cloud environment, what the challenges were from a feature engineering and modelling perspective and how they utilized a broad arsenal of techniques from supervised and unsupervised machine learning.

In this talk, Evgeniya discusses how multilingual CLIP-style models are changing the game in multimodal AI, particularly in Text-to-Image search applications. She will share her experience and techniques for fine-tuning multilingual and monolingual CLIP models on non-English data, specifically German data gathered through crowdsourcing. Additionally, she will demonstrate how to gather data in over 40 languages for fine-tuning such models. If you are interested in improving the performance of Text-to-Image search across multiple languages, this talk is for you!

The talk presents how ING has increased its productivity in delivering propensity-to-buy models for direct marketing campaigns by implementing an AutoML pipeline that utilizes the existing infrastructure. The implementation of clear standards and a highly automated model development process with state-of-the art features (SHAP, Bayesian parameter tuning, etc.) has decreased the time needed for development to deployment to less than one month while model updates only need a few hours.

Over the past years Schenker rolled out a multi-level AI training. Purpose of the training was to drive innovation culture and bring more data science savvy people in the organization to their next level. Currently, Schenker is redesigning the training into a broader-scoped data literacy training. This session shares some thoughts on what is really relevant to teach peers in an organization.

Machine learning models have become increasingly complex and resource-intensive, posing significant challenges. Model compression techniques addresses this issue by reducing the size and computational requirements of these models without significant loss in performance. This talk gives a comprehensive overview and code walk through of machine learning model compression techniques. (Model Pruning, Knowledge Distillation, Quantization, and low-rank approximation etc)

lexoffice allows small-medium businesses to step into a new era: cloud bookkeeping with an average booking automation rate of up to 70 %. In this talk, we will share key (data) ingredients for saving millions of hours for more than 250k customers and for scaling our daily work: experimentation setup, utilized metrics, insights to applied (ML) models, their validation and lifecycle in production. In addition, we will share the organizational evolution along this data product journey.

You are surrounded by fellow ML experts and Data Scientists and there is no better place to discuss and share your common problems. Off the record, and in a small group. These are your people – they understand your situation. Often rated the best part of the event, sharing your problems with like-minded professionals is your path to answers, a little empathy, and a stronger professional network. Come prepared with an issue you are facing, a problem you are solving, or a question that needs answering. And then, be ready and willing to help others to overcome their challenges.

Topics for the roundtable discussions:

From Data to Action: How can we achieve actual impact from our data (Philipp Freytag von Loringhoven)

In a world drenched with data, how do we navigate from raw information to real-world impact? In this discussion, we’ll dive into the complex maze of data transformation, exploring how it’s possible to catalyze tangible change from intangible information. Join a dynamic group of experts in analytics and business intelligence as they unlock the secrets behind successful data implementation in today’s challenging business environment.

ChatGPT Code Interpreter: Opportunity or Threat? (Peter Seeberg)

Do you, Data Scientist, use Code Interpreter for data cleaning, data merging, and analysis, including visualizations of your customer’s data? Have you become more productive since, by being handed initial explorational analysis and reducing repetitive tasks? Do you use Code Interpreter only for initial explorational analysis or also for more complex tasks? Do you, Domain Expert, use Code Interpreter for explorational analysis of your data? Have you become less dependent on your Data Scientist

Generative AI = Booster for Data-Driven Marketing & Sales? – How to leverage LLMs in the B2C Marketing Context (Cecilia Floridi)

Are LLMs just a hype or made to stay? How can we leverage the capabilities in the B2C marketing context, to improve the productivity and the value-add for consumers. We are going to have a open table discussion on use cases from practitioners and want to conclude which use cases worked out and why they succeeded or failed. The use cases shall incorporate various domains like data enrichment, text & content tagging, persona generation or the automation of the generation of highly personalized marketing campaigns. The focus should be on both the added value of the use cases and their operationalization.

Tired of collecting labels for your supervised Deep Learning model? Philip will show how they used a shortcut to create near infinite high quality labels to train a global water segmentation model on the VIIRS sensor in record time. Along the way, you will learn what is really important in an industry Machine Learning project, a lot about different satellite imagery and why you should not be a hero in choosing your model architecture.

Predictive modelers love building lots of models and then selecting the best model to deliver to the stakeholder. Popular metrics for model assessment and selection include R^2 and classification accuracy. Most often, however, the business doesn’t care about these and analysts may miss deploying the best models for the business. In this talk, a methodology for finding business-centric metrics will be explored so that the effectiveness of the models is significantly improved.

The objective of the Smartnet.ai project at Orange is the reduction of time between malfunction in the telecommunication system and recognition of its root cause. Although this project is for a Telco, its aim is universal for all industries dealing with constant service delivery. The Orange team put into production a scalable system on Google Cloud Platform. Its Deep Neural Network module allows for anomaly detection and data engineering for identifying the root cause. The system works in a streaming mode with data quality and drift monitoring modules.

Unfortunately, the majority of data projects fail. Yet, they fail for the same reasons. This talk will explore the most common reasons data projects fail and how to avoid them. We will do this by introducing the who, what, when, where, and how of data projects.